JP2015197805A - Processing system and multi-processing system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a technology of reducing latency of data transfer between a plurality of processing systems, when applications cooperate with each other in the processing systems.SOLUTION: A processing system 700 includes: a GPU 724 for generating a video; a memory 726 for storing data; a GPU data bus 718 for connecting the GPU 724 and the memory 726 to each other; a DMA controller 716 connecting to the GPU data bus 718; an input/output bridge 722 connecting to the GPU data bus 718, and connecting also to an external bus for transmitting/receiving data to/from an external processing system 700 which cooperates with the processing system 700; and a CPU 710 which shares data stored in a memory 726, with the GPU 724. The GPU 724 issues an instruction to the DMA controller 716 to transfer data stored in the memory 726, without using the CPU 710.

Description

  The present invention relates to a processing system and a multiprocessing system.

  In recent years, communication technologies for networks such as the Internet have rapidly developed, and so-called cloud services in which users are provided with various applications via networks have come to be provided. Among such cloud services, a cloud gaming service is also realized in which a game application is executed on a cloud server and an execution result is streamed and distributed to a client device via a network.

  On the other hand, recent advances in processor technology are remarkable, and not only the CPU (Central Processing Unit) responsible for general-purpose arithmetic processing, but also the performance of GPU (Graphics Processing Unit) mainly responsible for image processing. As the performance of these processors improves, the amount of computation required to generate video provided by the game application is also increasing.

  In order to realize a cloud gaming service, a plurality of game providing servers are required to provide different applications to a plurality of users. Of course, there are a plurality of processing systems for executing the game application.

  The inventor of the present application, in a server having a plurality of processing systems, allows a single application to be executed in cooperation with a plurality of processing systems, thereby achieving high image quality compared to a case in which an application is executed by a single processing system. Recognized the possibility of being able to provide services. At this time, in order to make collaborative work in a plurality of processing systems efficient, the necessity of reducing the latency of data transfer between the processing systems has been recognized.

  The present invention has been made in view of these problems, and an object of the present invention is to provide a technique for reducing latency of data transfer between processing systems when an application is executed in cooperation with a plurality of processing systems. is there.

  In order to solve the above-described problems, a processing system according to an aspect of the present invention includes a GPU that generates video, a memory that stores data, a data bus that connects the GPU and the memory, and a DMA ( Direct Memory Access) between the GPU and the input / output bridge that connects to the controller and the data bus and also connects to the external bus for sending and receiving data to and from the external processing system that cooperates with the processing system And a CPU that shares data stored in the memory. The GPU issues a data transfer instruction for data stored in the memory to the DMA controller without going through the CPU. When the DMA controller receives the data transfer instruction, the memory is sent to the input / output bridge to the external processing system. Transfer the data to be stored.

  Note that any combination of the above components and at least a part of the expression of the present invention converted between a method, an apparatus, a system, a computer program, a data structure, a recording medium, etc. are also effective as an aspect of the present invention. It is. The computer program may be provided as part of firmware incorporated in the device to perform basic control of hardware resources that implement the image processor. This firmware is stored in a semiconductor memory such as a ROM (Read Only Memory) or a flash memory in the device, for example. In order to provide the firmware or to update a part of the firmware, a computer-readable recording medium storing the program may be provided, and the program may be transmitted through a communication line.

  According to the present invention, it is possible to provide a technique that provides a technique for reducing the latency of data transfer between processing systems when an application is executed in cooperation by a plurality of processing systems.

It is a figure which shows typically the whole structure of the application provision system which concerns on embodiment. It is a figure which shows typically the internal structure of the blade server which concerns on embodiment. It is a figure which shows typically the circuit structure of the processing system which concerns on embodiment. It is a figure for demonstrating the data transfer between different processing systems. FIGS. 5A to 5C are diagrams schematically illustrating a connection topology between processing systems in one blade. It is a figure which shows typically the connection topology between the processing systems over a some braid | blade. It is a figure for demonstrating the memory map of a memory at the time of cooperation of four processing systems, and the dependence of a process. FIGS. 8A to 8E are diagrams illustrating an example of the flow of updating the memory map when the four processing systems cooperate. It is a flowchart which shows the flow of the change process of the allocation number of the processing system which the allocation server which concerns on embodiment performs.

  FIG. 1 is a diagram schematically showing an overall configuration of an application providing system 10 according to the embodiment. In FIG. 1, client devices 500 a and 500 b are illustrated among a plurality of client devices 500 collectively referred to as a client device 500. The client device 500 is connected to the login server 300 via a network 400 such as the Internet.

  The login server 300 manages an account of a user who uses the client device 500 connected via the network 400. The login server 300 refers to a database (not shown) that stores user names, passwords, and the like of users registered in the application providing system 10 and is responsible for user login processing and logoff processing.

  FIG. 1 illustrates blade servers 200a and 200b among a plurality of blade servers 200 collectively referred to as a blade server 200. The blade server 200 is a server that executes an application to be provided to the client device 500. The allocation server 100 allocates a blade server 200 that executes an application to be provided to the client device 500 from among the plurality of blade servers 200 and causes the application program to be executed. Details of the allocation process in the allocation server 100 will be described later.

  In most cases, the client device 500 has a calculation resource such as a CPU, and at least a part of the application program can be executed in a local environment of the client device 500. However, for the user who uses the client device 500, whether the application is executed with the local calculation resource or the calculation resource of the application providing system 10 on the network is not an important problem, and the service is used. It is the same in that it can be done. The client device 500 uses, for example, a PC (Personal Computer), a stationary game console, a portable game machine, a smartphone, a tablet PC, a fablet, an electronic book terminal, an electronic dictionary, a portable music player, a portable one-segment TV, and the like. realizable.

  FIG. 1 illustrates a state in which a client device 500 and a login server 300 are connected via a network 400, and schematically illustrates an actual system configuration. However, a user who uses the client device 500 does not often recognize the system configuration of the login server 300 and the allocation server 100 including the network 400. From the standpoint of the user who uses the application providing system 10, it is not necessary to understand the physical configuration or software configuration of the application providing system 10 in order to receive the service, and it is as if in the cloud of the network. I get the impression that I use the services that came out of. An application providing system 10 shown in FIG. 1 constitutes a cloud gaming system.

  Hereinafter, in the application providing system 10 according to the embodiment, a case where the blade server 200 executes game content developed on the premise that the blade server 200 is executed on a game console that is dedicated hardware will be described. In this sense, the blade server 200 functions as a game providing server. However, the content executed by the blade server 200 is not limited to a game, and the present embodiment also includes a case where content of a genre other than a game such as an SNS (Social Networking Service) or a video browsing service is executed.

  FIG. 2 is a diagram schematically illustrating an internal configuration of the blade server 200 according to the embodiment. As shown in FIG. 2, one blade server 200 can mount 16 blades 600 in a housing. The blade server 200 according to the embodiment includes 16 blades 600. In FIG. 2, 16 blades 600 are illustrated as blades 600a to 600p. Hereinafter, unless otherwise distinguished, the plurality of blades 600 are simply referred to as “blades 600”. The blade 600 alone constitutes an application providing server, and 16 blades 600 are not necessarily required to execute the application program. Further, the number of blades 600 included in the blade server 200 according to the embodiment is not limited to 16, and it is only necessary to include a plurality of blades 600.

  As shown in FIG. 2, one blade 600 includes four processing systems 700a to 700d (hereinafter simply referred to as “processing system 700” unless otherwise specified), a power port 610, and an input / output. A port 620 and a communication port 630 are provided on the same substrate.

  As described above, the blade server 200 according to the embodiment executes game content developed on the premise that the blade server 200 is executed on a game console that is dedicated hardware. The four processing systems 700a to 700d arranged on the substrate constituting the blade 600 are each a processing system having performance equivalent to that of the processing system included in the game console.

  In general, game titles developed for dedicated game consoles are highly optimized for hardware resources such as processors and memories. For this reason, one processor is occupied to execute one game title, and it is difficult for a single processor to execute a plurality of game titles simultaneously. The processing system 700 according to the embodiment is also a game processing system that can execute at most one game application alone. The blade 600 according to the embodiment includes a plurality of such processing systems 700 and constitutes a multiprocessing system.

  The power port 610 supplies power to the processing system 700 included in the blade 600. The input / output port 620 is a port for transmitting / receiving data to / from another blade 600, and can be realized by using a known input / output interface such as PCI Express, for example. The communication port 630 is a port that receives an allocation instruction of the processing system 700 that executes the game application from the allocation server 100. The communication port 630 can be realized by using a LAN port conforming to a known local area network (LAN) standard such as Ethernet (registered trademark).

  FIG. 3 is a diagram schematically illustrating a circuit configuration of the processing system 700 according to the embodiment. A processing system 700 according to the embodiment includes a CPU 710, a GPU 724, a memory 726, a memory control unit 714, a bridge 712, a DMA controller 716, an input / output bridge 722, a GPU bus 718, and a CPU bus 720.

  The GPU 724 generates a video image of the game application provided to the client device 500. The GPU 724 is connected to the memory 726 via the GPU bus 718. The memory 726 stores the data output from the GPU 724.

  The CPU 710 is connected to the memory 726 via the CPU bus 720. The CPU 710 is also connected to the GPU 724 via the GPU bus 718. The CPU 710 is connected to these buses via the bridge 12. The memory 726 is connected to the GPU bus 718 and the CPU bus 720. For this reason, the exchange of data with the CPU 710 or the GPU 724 in the memory 726 is executed via the memory control unit 714. The CPU 710 and the GPU 724 share the memory 726 and share data via the memory 726. Both the GPU bus 718 and the CPU bus 720 function as data buses used for data exchange.

  The DMA controller 716 is connected to the GPU bus 718. The input / output bridge 722 is connected to the GPU bus 718 and serves as an interface for transmitting / receiving data to / from the external processing system 700. The DMA controller 716 is connected to the GPU bus 718, and can transfer data stored in the memory 726 to the external processing system 700 under the control of the GPU 724 without going through the CPU 710.

  FIG. 4 is a diagram for explaining data transfer between different processing systems 700. In the example shown in FIG. 4, two different processing systems 700a and 700b are connected via input / output bridges 722a and 722b. More specifically, the input / output bridge 722a included in the processing system 700a includes a plurality of input / output ports 730a, 731a, 732a, and 733a for transmitting / receiving data to / from other processing systems 700. Similarly, 730b included in the processing system 700b includes a plurality of input / output ports 730b, 731b, 732b, and 733b. The processing systems 700a and 700b are connected via an input / output port 730a and an input / output port 730b.

  In FIG. 4, in addition to the processing system 700b, the processing system 700a is connected to the processing system 700c via the input / output port 732a, and is also connected to the processing system 700d via the input / output port 733a. In addition to the processing system 700a, the processing system 700b is also connected to the processing systems 700c, 700d, and 700e via input / output ports 731b, 732, and 733b, respectively. As described above, the bus connecting different processing systems 700 via the input / output ports functions as an external bus for transmitting and receiving data between the processing systems 700.

  The input / output bridge 722 maps not only the access via the DMA controller 716 but also each input / output port to a specific range of physical memory, and sends a command that flows to the GPU bus 718 to a remote node (another processing system 700). It can also be sent to the GPU bus 718. The input / output bridge 722 is connected to the GPU bus 718 and is mapped to physical memory. Therefore, each of the CPU 710 and the GPU 724 can directly access the memory 726 of the remote node.

  Hereinafter, a case where the processing system 700a and the processing system 700b execute one game application program in cooperation will be described as an example.

  In the processing system 700 according to the embodiment, the GPU 724 can issue a data transfer instruction for data stored in the memory 726 to the DMA controller 716 without being controlled by the CPU 710. For example, it is assumed that the GPU 724a issues a data transfer instruction that instructs the DMA controller 716a to transfer the data stored in the memory 726a to the processing system 700b. As a trigger for issuing the data transfer instruction, the GPU 724a may issue a data transfer instruction spontaneously. Alternatively, the data transfer command may be issued in response to the data read command broadcast by the processing system 700b. In any case, when receiving the data transfer command, the DMA controller 716a causes the input / output bridge 722 to transfer the data stored in the memory 726 to the processing system 700b.

  The DMA controller 716b in the processing system 700b receives the data transferred from the processing system 700a via the input / output bridge 722b. Subsequently, the DMA controller 716b writes the received data to the memory 726b in the processing system 700b.

  When the GPU 724b in the processing system 700b acquires data stored in the memory 726a in the processing system 700a, first, the GPU 724b issues a data acquisition command to the DMA controller 716b without going through the CPU 710b. When receiving the data acquisition command, the DMA controller 716b issues a data read command to the input / output bridge 722b. The input / output bridge 722b broadcasts a data read command issued by the DMA controller 716b to the external bus. The operation after the processing system 700a responds to this command is the same as described above.

  As described above, the GPU 724 can transmit the data stored in the memory 726 to the external processing system 700, or can receive the data of the external processing system 700 and store it in the memory 726. Details of processing while sharing data among the plurality of processing systems 700 will be described later.

  FIGS. 5A to 5C are diagrams schematically illustrating a connection topology of the processing system 700 in one blade 600. FIG. As described above, the processing systems 700 a to 700 d having the same performance are mounted on the substrate constituting the blade 600. Here, it is assumed that the input / output bridges 722a to 722d in each processing system 700 are mounted using a known PCI Express.

  Although not shown, the blade 600 is provided with a system controller that controls the operation of PCI Express. The system controller receives a command for instructing a connection relationship from the allocation server 100 via the communication port 630 and controls the connection of the input / output bridge 722. More specifically, the system controller sets Root Complex and Endpoint in PCI Express based on the received command, and then establishes a connection by applying a reset. The processing systems 700 that have established connections also share the data stored in the memory, and can execute the same game application in cooperation.

  FIG. 5A shows a case where two processing systems 700a and 700b are connected, and FIG. 5B shows a case where three processing systems 700a, 700b, and 700c are connected. FIG. FIG. 5C is a diagram illustrating a case where four processing systems 700a, 700b, 700c, and 700d are connected. In each processing system 700, the input / output bridge 722 includes a plurality of input / output ports (730, 731, 732, and 733), and in any case, the processing systems 700 are point-to-point (point-to-point). point). As a result, the processing systems 700 can communicate one-on-one.

  FIG. 6 is a diagram schematically showing a connection topology between the processing systems 700 across a plurality of blades 600. FIG. 6 shows an example in which four blades 600a, 600b, 600c, and 600d are connected. The four processing systems 700 in each blade 600 are connected to each other in a point-to-point manner as in the case shown in FIG.

  The four blades 600a, 600b, 600c, and 600d are connected to communicate one-on-one via an input / output bridge 722 included in any processing system 700 in each blade. For example, the blade 600a and the blade 600d are connected to each other in a point-to-point manner via the input / output bridge 722 of the processing system 700d in the blade 600a and the input / output bridge 722 of the processing system 700d in the blade 600d. . The processing system 700d is also connected to the processing systems 700a, 700b, and 700c in the blade 600a, but the input / output bridge 722 includes four input / output ports 730, 731, 732, and 733, so that the four different processing systems 700 and Can be connected at the same time.

  When 16 processing systems 700a to 700p are connected in the configuration shown in FIG. 6, any different processing systems 700 are connected to each other at a maximum of 3 hops, that is, processing that passes through until reaching the other party. The maximum number of systems 700 is three. Thereby, when 16 processing systems 700a-700p cooperate and execute one game application, the latency at the time of data transfer at the time of data sharing can be suppressed.

  In general, when a plurality of GPUs are operated in cooperation, the GPU drawing processing is divided into frame units or region units, and each drawing task is distributed to a plurality of GPUs. The mainstream is that each GPU processes a plurality of distributed drawing tasks and writes the drawing result back to one GPU. This is because there are restrictions on the transmission of drawing results between GPUs.

  For example, when a dedicated bus is used for transmitting drawing results, the dedicated bus is like a display interface optimized for image transmission, and thus the bandwidth is guaranteed. However, the bandwidth can be limited to a bandwidth necessary for transmitting an image having a certain resolution or less. On the other hand, PCI Express may be used for transmission of drawing results. In this case, the bandwidth is large and no dedicated interface is required, but it is difficult to guarantee the bandwidth because it is mixed with other data transmissions.

  In recent years, GPUs have been able to perform not only graphics rendering but also parallel processing on behalf of the CPU for a part of physical computations and simulations that have been performed by the CPU. For this reason, in real-time applications such as games, it has become possible to divide physics and the like into processes suitable for the CPU and GPU, and to reflect the results in the drawing process after executing them in parallel. However, for example, when reflecting user input in physical calculation and drawing, distributed processing in units of frames is a factor that increases latency, which is not suitable.

  In addition, when physics calculation and drawing are completed within one frame, the CPU and GPU interaction occur multiple times during one frame. For example, consider a case in which one frame is divided into four parts of 2 in the vertical direction and 2 in the horizontal direction, and each GPU draws each area. At this time, the shadow of a moving object existing in a region processed by a certain GPU may be reflected not only in that region but also in a region processed by another GPU. In this case, the movement of the moving object is processed by the CPU, and two different GPUs draw the shadow caused by the moving object, and the interaction between the CPU and the GPU occurs a plurality of times during one frame.

  A processing system 700 according to the embodiment is a system in which a CPU 710 and a GPU 724 are integrated into one chip. Further, the CPU 710 and the GPU 724 share the memory 726. For this reason, when the CPU 710 and the GPU 724 refer to common data, no memory copy occurs and the latency is shortened. Further, since the DMA controller 716 is connected to the GPU bus 718 and can perform data transfer under the control of the GPU 724, it is synchronized with communication even when distributed processing is performed with other processing systems 700 connected via the input / output port 732. The latency due to can be suppressed. Accordingly, real-time distributed processing in which data is exchanged between different GPUs 724 and CPUs 710 multiple times within one frame can be realized.

  FIG. 7 is a diagram for explaining the memory map of the memory 726 and the processing dependency when the four processing systems 700a, 700b, 700c, and 700d cooperate. In FIG. 7, the symbol “a” indicates a memory map of the memory 726 of the processing system 700a. A memory area indicated by symbol A (hereinafter referred to as “area A”) indicates an area for storing data read from a storage (not shown) by the CPU 710 of the processing system 700a.

  A memory area indicated by a symbol B (hereinafter referred to as “area B”) indicates a memory area in which the result of arithmetic processing by the GPU 724 using the data of area A is stored. An area indicated by a symbol C (hereinafter referred to as “area C”) indicates a memory area in which the result of arithmetic processing performed by the GPU 724 using both the data in the area A and the data in the area B is stored. An area indicated by a symbol D (hereinafter referred to as “area D”) is a memory area for storing the result of arithmetic processing performed by the CPU 710 using data after the data in the area C is completely prepared. An area indicated by a symbol FB (hereinafter referred to as “area FB”) indicates a frame buffer (Frame Buffer) that stores the results of drawing using the data of area A, the data of area B, and the data of area D.

  FIGS. 8A to 8E are diagrams illustrating an example of the flow of updating the memory map when the four processing systems 700a, 700b, 700c, and 700d cooperate. The memory map is updated in the order of FIG. 8A to FIG.

  In FIG. 8 (a)-(e), the symbols “a”, “b”, “c”, and “d” are memory maps of the memory 726 of the processing systems 700a, 700b, 700c, and 700d, respectively. Indicates. The memory areas indicated by the symbols “A”, “B”, “C”, “D”, and “FB” are the same as described above.

  FIG. 8 illustrates an example in which the results of distributed processing performed by the four processing systems 700a, 700b, 700c, and 700d are collected in the processing system 700d. The game program and data for executing the distributed processing are stored in a shared storage (not shown) accessible from all processing systems 700. Execution of the program is executed by each processing system 700 being read from the shared storage, but after the huge data is read from the processing system 700d, it is broadcast to all other processing systems 700 via the input / output bridge 722d. The As for the data to be shared, an area A reserved in advance is used so that all the processing systems 700 share the data at the same address.

  The processing system 700d reads all data necessary for program execution, such as textures and primitives, from the shared storage. The processing system 700d stores the read data in the area A. The processing system 700d continues to read data from the shared storage, and broadcasts the contents of the area A to the other processing systems 700a, 700b, and 700c as low-priority uncompressed data. As a result, as shown in FIG. 8A, the four processing systems 700a, 700b, 700c, and 700d store data common to the respective areas A.

  When the four processing systems 700a, 700b, 700c, and 700d store common data in their respective areas A, the CPU 710 of the processing system 700d transmits a message for informing the start of program execution to the other processing systems 700a, 700b, And 700c with high priority. In the processing systems 700a, 700b, and 700c, tasks to be executed by the GPU 724 have already been executed, and each GPU 724 polls a message area (not shown).

  When each GPU 724 detects the message, it uses the data stored in the area A in each memory 726 and outputs the result to the area B in each memory 726. FIG. 8A shows that each of the four processing systems 700 a, 700 b, 700 c, and 700 d stores the calculation result in the respective region B.

  When each GPU 724 outputs the result to the area B, the next command has already been input to each GPU 724. The next command input to the GPU 724 is an instruction for broadcasting the contents of the area B to the area B in the memory 726 in the other processing system 700. Each DMA controller 716 broadcasts the contents of region B to other processing systems 700 in accordance with the DMA instruction issued by GPU 724. Thereby, each processing system 700 can share the data of the area B processed by itself with other processing systems 700. FIG. 8B shows a result of further processing the unprocessed data after each processing system 700 shares the data in area B.

  Here, the DMA controller 716 according to the embodiment compresses the data stored in the memory 726, decompresses the data received from the data transferred from the external processing system 700 by the input / output bridge 722, and issued by the GPU 724 or the CPU 710. Perform instruction priority control. More specifically, when a compression option is set in the received write command, the DMA controller 716 transfers data read from the memory 726 to a compression engine (not shown). In order to conceal the latency generated by the compression engine, the DMA controller 716 processes requests input to other command interfaces in parallel.

  When receiving the compressed data from the input / output bridge 722, the DMA controller 716 transfers the received data to a decompression engine (not shown). The decompression engine writes the decompressed data in the memory 726. When the destination of the read / write command received from the input / output bridge 722 is the local node, the DMA controller 716 sends the command to the GPU bus 718. If the destination is not a local node, the DMA controller 716 processes the instruction as an error or routes it to an input / output port connected to the destination node.

  As described above, by compressing data when transferring data between different processing systems 700, it is possible to improve transfer efficiency.

  Returning to the description of FIG. As described above, the area C is a memory area that stores the result of calculation processing performed by the GPU 724 using both the data in the area A and the data in the area B. The GPU task that generates the data to be stored in the area C is polling the data stored in the area B and waits until the data necessary for the operation is available. FIG. 8B shows that the data for the area C is generated and stored in each processing system 700 because the data for generating the data to be stored in the area C is ready. FIG. 8C shows a case where the data in the area C generated in each processing system 700 is broadcast to each other, and all the processing systems 700 share the data in the area C.

  Here, data to be written to the area B from another processing system 700 is written from the DMA controller 716 via the GPU bus 718. On the other hand, data to be written in the area C is written via the CPU bus 720. In addition, when data is written to another processing system 700, an instruction may be issued to the DMA controller 716 so as to write to the memory via the CPU bus 720.

  When the CPU 710 of the processing system 700d receives the write completion interrupt for the area C, the CPU 710 generates data to be stored in the area D using the data of the area C as shown in FIG. Note that since the writing of the area C is performed via the CPU bus 720, even if the CPU 710 caches the contents of the area C, the contents of the cache are updated by external writing. When the CPU 710 of the processing system 700d stores data in the area D, the CPU 710 broadcasts the data to other processing systems 700. FIG. 8D shows a case where the data in area D is shared by all the processing systems 700.

  The drawing task processed by the GPU 724 in each processing system 700 polls the writing completion of the area C and the area D, and starts the frame generation process when the data is ready. The rendering result of each GPU 724 is written in the respective area FB. FIG. 8D shows a case where the rendering result of each GPU 724 is written in each area FB. The drawing result of each GPU 724 is broadcast to the area FB of the processing system 700d, and is copied into the area FB of the processing system 700d as shown in FIG. When the data in the area FB is aligned, the processing system 700d performs display output.

  As described above, the blade server 200 according to the embodiment includes a plurality of processing systems 700 that can execute at most one game application as a single unit, and executes one game application in cooperation with the plurality of processing systems 700. be able to. At this time, the GPU 724 in each processing system 700 can exchange not only final processing results but also intermediate data closely. Thereby, it is possible to execute a process with a larger amount of calculation than when the game application is executed by one processing system 700, and to provide a more precise and high-definition image.

  When PCE Express 4.0 (Gen4) is adopted as the input / output bridge 722, it is assumed that 16 lanes are divided into 4 ports to be 4 [lane / port]. At this time, the bandwidth per port is 8 GB / second, and the effective bandwidth is about 5 GB / second. The peak effective bandwidth when data is simultaneously exchanged from 4 ports is 20 GB / second for each TX / RX. When combined with lossy compression, the internal bandwidth is several times that.

  Here, the amount of data that can be sent in 1 millisecond with a bandwidth of 5 GB / second is 5 MB. When converted into 32-bit vertex data, it becomes 1.25 million vertices per millisecond. Therefore, when texture data of 32 bits / pixel is compressed to about 1/10, 50 MB per millisecond, that is, data of 12.5 textures of an image of 1024 pixels vertically and 1024 pixels horizontally can be transferred.

  As described above, according to the blade server 200 according to the embodiment, each processing system 700 is connected point-to-point, the priority control of processing instructions in the DMA controller 716, and the GPU that is an internal bus With the bandwidth guarantee by the bus 718, it is possible to guarantee the bandwidth of the entire system. For example, processing huge amounts of data read from the file system in the background while each GPU 724 processes real-time data of a determined size delays the processing that must be completed within one frame. Can be suppressed.

  In each processing system 700, the CPU 710 and the GPU 724 can reduce the latency of data transfer between the processing systems 700 by performing a memory copy via the DMA controller 716. This enables distributed processing in which processing by the CPU 710 and GPU 724 is mixed while exchanging data multiple times within one frame, which has been difficult in the past.

  The case where a plurality of processing systems 700 cooperate to execute one game application has been described above. Next, the allocation process of the processing system that is responsible for executing one game application will be described.

  In a cloud gaming system, when providing a high quality and high load game, it is difficult to move a plurality of games with a single processing system. Therefore, a plurality of blade servers 200 including a plurality of processing systems 700 are prepared as in the application providing system 10 illustrated in FIG. However, if the processing system 700 is prepared so that it can cope with the peak time when the load on the system is the largest, the load decreases depending on the time zone, and the number of processing systems 700 that do not operate increases. When the application providing system 10 is operated as a game providing system, it is necessary to keep latency low, and it is difficult to realize load distribution in a plurality of regions having different time zones.

  Therefore, the allocation server 100 according to the embodiment sets the number of processing systems 700 to be allocated to execute one game application based on the load of the blade server 200 that is a game providing server. More specifically, the allocation server 100 sets a smaller number of processing systems 700 to be allocated to execute one game application when the load on the blade server 200 is large than when the blade server 200 is small.

  The allocation server 100 may also increase the operation clocks of the GPU 724 and the CPU 710 in the processing system 700 allocated to execute the game application when the load of the blade server 200 is small, as compared to when the blade server 200 is large. The allocation server 100 switches the operating state of the processing system 700 as follows, for example, according to the load of the blade server 200.

State 0: Standby State 1: One processing system 700 is used with a normal clock State 2: One processing system 700 is used with a clock up State 3: Four processing systems 700 are used with a clock up State 4:16 Use processing system 700 with clock up

  The allocation server 100 sets the operating state of the processing system 700 to state 4 when the load of the blade server 200 is small, and changes the operating state to state 3, state 2, and state 1 as the load increases. Thereby, when the load of the blade server 200 is small, it is possible to provide a game to the client device 500 with a fine and high-definition image. Even if the load on the blade server 200 increases, it is possible to provide a game to the client device 500 while maintaining the same image quality as when executed on a game console, which is dedicated hardware. Note that the above-described state is an example, and in addition to this, a state of a different combination of the number of processing systems 700 and clocks can be defined.

  Here, the “load of the blade server 200” is an allocation index of the processing system 700 that is used by the allocation server 100 to determine the number of processing systems 700 that execute one game application. Specifically, it is at least one of the power consumption of the blade server 200, the exhaust heat amount of the blade server 200, and the number of client devices 500 or users who are providing the game application to the blade server 200. The power consumption and the amount of exhaust heat of the blade server 200 can be acquired using a power meter or a temperature sensor (not shown). When the power consumption of the blade server 200 is large, when the amount of exhaust heat is large, or when the number of client devices 500 or users providing the game application is large, the allocation server 100 is compared with the case where the blade server 200 is not. It is determined that the load of 200 is large.

  When the allocation server 100 is in the operating state of the processing system 700, the load of the blade server 200 may fluctuate, for example, when the user ends the game or switches to a lightly loaded application. . Therefore, when the load on the blade server 200 increases or decreases while providing a game application to a plurality of client devices 500, the allocation server 100 receives the game application triggered by the scene switching in the game application that is being provided. The number of processing systems 700 assigned to perform

  Here, “switching of scenes in a game application” refers to, for example, redrawing the entire frame by switching the stage according to the progress of the game or changing the position of a virtual viewpoint for composing the video of the game. Means timing. When the allocation server 100 changes the operating state of the processing system 700, the image quality of the game application to be provided is also changed. However, by changing the operating state at the time of switching scenes, it is possible to effectively use the resources of the application providing system 10 while suppressing the uncomfortable feeling that the user feels as the image quality changes.

  As described above, even when a game application is provided in a state 1 in which one processing system 700 is used with a normal clock, the client device 500 has an image quality equivalent to that executed on a game console that is dedicated hardware. Can be provided. Therefore, when the game application is provided in other operating states, the game application can be provided with higher image quality than when the game application is executed on a dedicated game console.

  Accordingly, when the assignment server 100 assigns a plurality of processing systems 700 to execute the game application being provided to the client device 500, the assignment server 100 provides the client device 500 with a high quality game application. You may be notified. This can be realized, for example, by displaying a message indicating the “high image quality mode” on a part of the video provided to the client device 500, or by giving a voice when starting to provide the game. Thereby, since the user can recognize that he / she is playing a game with higher image quality than when the game console is executed at home, it can be an incentive to use the application providing system 10.

  FIG. 9 is a flowchart showing a flow of processing for changing the number of allocations of the processing system 700 executed by the allocation server 100 according to the embodiment.

  The allocation server 100 confirms the load of the blade server 200 in the application providing system 10 (S2). When the load of the blade server 200 is increasing (Y in S4), the allocation server 100 decreases the setting of the allocation number of the processing system 700 allocated to the execution of one game application (S6).

  When the load on the blade server 200 has not increased (N in S4) and the load on the blade server 200 has decreased (Y in S8), the allocation server 100 allocates processing for execution of one game application. The setting of the allocation number of 700 is increased (S10). After the allocation server 100 changes the setting of the allocation number of the processing system 700, if there is a scene change in the game application being executed (Y in S12), the allocation server 100 reflects the setting of the allocation number (S14). While there is no scene change in the game application being executed (N in S12), the allocation server 100 waits for the setting number allocation to be reflected until there is a scene change.

  If the allocation server 100 reflects the number of settings or the load on the blade server 200 does not increase or decrease (N in S8), the processing in this flowchart ends.

  As described above, according to the assignment server 100 according to the embodiment, the resources of the application providing system 10 can be adaptively assigned according to the load of the blade server 200 that is a game providing server.

  The present invention has been described based on the embodiments. The embodiments are exemplifications, and it will be understood by those skilled in the art that various modifications can be made to combinations of the respective constituent elements and processing processes, and such modifications are within the scope of the present invention. .

[Modification]
In the above description, the allocation server 100 changes the allocation number of the processing system 700 allocated for execution of one game application according to the load of the blade server 200, or changes the operation clock of the processing system 700. Explained the case. The assignment server 100 may change the number of assignments and the number of clocks in accordance with the client device 500 that provides the game application, instead of or in addition to the load of the blade server 200.

  For example, when a user registers in the application providing system 10 in order to use the application providing system 10 according to the embodiment, a case where the user pays a fee according to a usage mode of the application providing system 10 is considered. For example, a user can register as a “premium user” to whom the processing system 700 is assigned in preference to a normal user by paying a higher fee. When the user who uses a certain client device 500 is a premium user, the allocation server 100 is in the state 2 or the state when providing a game application to the client device 500 even when the load on the blade server 200 is large. Assign a better working state. Thereby, it is possible to provide a specific user with a game application with a better image quality than when executing on a game console that is always dedicated hardware.

  The invention according to the present embodiment may be specified by the items described below.

[Item 1-1]
A processing system,
GPU (Graphics Processing Unit) that generates video,
Memory to store data,
A data bus connecting the GPU and the memory;
A DMA (Direct Memory Access) controller connected to the data bus;
An input / output bridge connected to the data bus and also connected to an external bus for transmitting and receiving data to and from an external processing system cooperating with the processing system;
A CPU (Central Processing Unit) that shares data stored in the memory with the GPU;
The GPU issues a data transfer instruction of data stored in the memory to the DMA controller without going through the CPU,
The DMA controller, when receiving the data transfer command, causes the input / output bridge to transfer data stored in the memory to the external processing system.
[Item 1-2]
The GPU issues a data acquisition command to the DMA controller without going through the CPU,
When the DMA controller receives the data acquisition command, it issues a data read command to the input / output bridge,
The input / output bridge broadcasts a data read command issued by the DMA controller to the external bus and receives data transferred from the external processing system in response to the read command,
The processing system according to item 1-1, wherein the DMA controller writes data received by the input / output bridge to the memory.
[Item 1-3]
Item 1-2 is characterized in that the DMA controller controls compression of data stored in the memory and expansion of data received from the data transferred by the input / output bridge from the external processing system. The processing system described.
[Item 1-4]
A multiprocessing system comprising two or more processing systems according to any one of items 1-1 to 1-3,
Each input / output bridge in each processing system has two or more input / output ports for transmitting and receiving data to and from other processing systems.
Each of the processing systems is connected to be capable of one-to-one communication through an input / output port of an input / output bridge in each game processing system.
[Item 1-5]
When the GPU or CPU in each processing system processes and stores data in each memory, the GPU or CPU that processed the data issues a data transfer instruction to the DMA controller,
Item 5. The multiprocessing system according to item 1-4, wherein the DMA controller that has acquired the data transfer command broadcasts the data stored in the memory to each processing system via an input / output bridge.
[Item 1-6]
A plurality of blades in which a plurality of the processing systems according to any one of Items 1-1 to 1-3 are arranged on the same substrate;
A plurality of game processing systems in each blade are connected to communicate one-on-one via input / output ports of the input / output bridge,
Each of the plurality of blades is connected so as to be capable of one-to-one communication via an input / output port of an input / output bridge included in any of the game processing systems in each blade. Item 6. A multiprocessing system according to item 1-5.

[Item 2-1]
An allocation server connected to a plurality of client devices via a network;
A game providing server that executes a game application provided to the client device assigned by the assignment server,
The game providing server is a multi-processing system having a plurality of game processing systems capable of executing a maximum of one game application alone,
The allocation server sets a number of game processing systems to be allocated to execute one game application in the game providing server based on a load of the game providing server.
[Item 2-2]
The allocation server sets a smaller number of game processing systems to be allocated to execute one game application in the game providing server when the load of the game providing server is large than when the game providing server is small. The game providing system according to item 2-1.
[Item 2-3]
Each of the game processing systems in the game providing server is
A GPU for generating video of the game application;
A memory for storing data generated by the GPU;
A CPU that shares data stored in the memory with the GPU;
A DMA controller (Direct Memory Access System) that transmits and receives data stored in the memory to and from other game processing systems;
Item 2-1 or Item 2-2 is characterized in that the allocation server increases the operation clock of the GPU and the CPU when the load of the game providing server is small compared to when the load is large. Game provision system.
[Item 2-4]
The load of the game providing server is at least one of the power consumption of the game providing server, the amount of heat exhausted from the game providing server, and the number of client devices that the game providing server is providing game applications. The game providing system according to any one of items 2-1 to 2-3, which is characterized.
[Item 2-5]
When the load of the game providing server increases or decreases while providing a game application to a plurality of client devices, the allocation server executes the game application in response to a scene switching in the game application being provided 5. The game providing system according to any one of items 2-1 to 2-4, wherein the number of game processing systems to be allocated is increased or decreased.
[Item 2-6]
When the allocation server allocates a plurality of game processing systems to execute a game application being provided to the client device, the allocation server notifies the client device that the game application is provided with high quality. The game providing system according to any one of items 2-1 to 2-5, characterized by:

  10 application providing system, 12 bridge, 50 texture, 100 allocation server, 200 blade server, 300 login server, 400 network, 500 client device, 600 blade, 610 power supply port, 620 input / output port, 630 communication port, 700 processing system, 710 CPU, 712 bridge, 714 memory controller, 716 DMA controller, 718 GPU bus, 720 CPU bus, 722 I / O bridge, 724 GPU, 726 memory, 730, 732 I / O port.

Claims (6)

A processing system,
GPU (Graphics Processing Unit) that generates video,
Memory to store data,
A data bus connecting the GPU and the memory;
A DMA (Direct Memory Access) controller connected to the data bus;
An input / output bridge connected to the data bus and also connected to an external bus for transmitting and receiving data to and from an external processing system cooperating with the processing system;
A CPU (Central Processing Unit) that shares data stored in the memory with the GPU;
The GPU issues a data transfer instruction of data stored in the memory to the DMA controller without going through the CPU,
The DMA controller, when receiving the data transfer command, causes the input / output bridge to transfer data stored in the memory to the external processing system. The GPU issues a data acquisition command to the DMA controller without going through the CPU,
When the DMA controller receives the data acquisition command, it issues a data read command to the input / output bridge,
The input / output bridge broadcasts a data read command issued by the DMA controller to the external bus and receives data transferred from the external processing system in response to the read command,
The processing system according to claim 1, wherein the DMA controller writes data received by the input / output bridge to the memory.   3. The DMA controller according to claim 2, wherein the DMA controller controls compression of data stored in the memory, and expansion of data received from data transferred from the external processing system by the input / output bridge. Processing system. A multi-processing system comprising two or more processing systems according to any one of claims 1 to 3,
Each input / output bridge in each processing system has two or more input / output ports for transmitting and receiving data to and from other processing systems.
Each of the processing systems is connected to be capable of one-to-one communication through an input / output port of an input / output bridge in each game processing system. When the GPU or CPU in each processing system processes and stores data in each memory, the GPU or CPU that processed the data issues a data transfer instruction to the DMA controller,
5. The multiprocessing system according to claim 4, wherein the DMA controller that has acquired the data transfer instruction broadcasts the data stored in the memory to each processing system via an input / output bridge. A plurality of blades in which a plurality of the processing systems according to any one of claims 1 to 3 are arranged on the same substrate,
A plurality of game processing systems in each blade are connected to communicate one-on-one via input / output ports of the input / output bridge,
6. The plurality of blades are connected so as to be capable of one-to-one communication via an input / output port of an input / output bridge provided in any one of the game processing systems in each blade. The multiprocessing system described in

Images